Definitely, not the typical questions you get asked in a PyCon conference; but that’s the kind of questions that you get asked when you are a hybrid quant / data scientist and a Keynote speaker (PyCon Asia & the Pacific 2017).
In finance, is domain expertise more important than machine learning skills?I am a firm believer that the application of data science, machine learning and AI without domain expertise is a serious handicap for the companies and organizations starting to use these tools (at least in finance) and hiring talent straight out of academia, w/o exposure to real life problems.
A domain expert is someone who is not necessarily related to the technology aspect but has in-depth knowledge about the particular industry, how it is shaping up, the trends, and what are the things that will impact the industry.
Don’t get me wrong, I love how easy it is to set up models now with, for example TensorFlow 2.
0 with Keras (yes, now TF is full Keras, read Cassie Kozyrkov ’s post here, can’t wait to start playing with the Alpha version), versus developing neural nets from scratch like I did for the hedge fund I first worked for in 1993.
But if you know the tools, having domain expertise still gives you a tremendous edge over your plain vanilla data scientist.
Not all data scientists / quants are equal: the different types of Data ScientistsIn the Strata Conference in 2012 in Santa Clara, California which I had the opportunity to attend, there was a “Data Science Debate” among industry experts and the consensus then was that in data science, domain expertise was more important than machine learning skills.
Nevertheless, quick iteration reduces the advantage of having domain expertise, but domain expertise and quick iteration is unbeatable, particularly in a financial environment when you need to act fast.
I have seen some top Silicon Valley data scientists fail in finance and insurance problems, where a mid-level Wall Street quant would have done just fine.
In my current consulting practice, I have seen companies with very specific objectives matching the wrong type of data scientist to the role’s expectations.
Cassie Kozyrkov: “Hope is not a strategy.
Just hoping that you hire a researcher and they transcends into the applied stuff.
ah.
that’s a very risky proposition there”In finance there are basically 4 types of quants: algorithmic trading quants, front office quants, researcher quants, and developer quant.
In data science there are also 4 basic types: business data scientist, creative data scientist, developer data scientist and researcher data scientist.
The common misconception is that the “researcher data scientist” (like the “researcher quant” typically, a PhD with impressive academic credentials but little practical expertise) can do it all in a business setting, when in reality, team work with other types of data scientists is necessary.
There is an excellent video I watched recently, where Cassie Kozyrkov talks about this issue.
As a matter of fact, a PhD “researcher quant” might have unknowingly been the largest, single contributor to the financial crisis, by means of algos he developed that rapidly spread across the whole industry since it facilitated business.
What is Alpha?Since the title of this article states alpha generation, let’s define alpha and other terms for the benefit of data scientists reading this with zero domain expertise in finance.
Alpha is a measure of the active return on an investment, the performance of that investment compared with a benchmark (the market).
An alpha of 1% means the investment’s return on investment over a selected period of time was 1% better than the market during that same period; a negative alpha means the investment underperformed the market.
Alpha, along with beta, is one of two key coefficients in the capital asset pricing model used in modern portfolio theory and is closely related to other important measures such as standard deviation, R-squared and the Sharpe ratio.
Historically, the vast majority of traditional funds have had negative alphas, which has led to a flight of capital to index funds and non-traditional hedge funds.
(Why would you pay fees to somebody who can’t even beat the market?)With that explained, let’s dive into some background information.
One-of deals: the Term Asset Loan Facility (TALF)Merriam Webster defines “one-off” as “limited to a single time, occasion, or instance : “one-shot”.
The TALF was one of such one-off deals, created by the US Federal Reserve to mitigate the problems of the financial crisis.
The TALF program was created in November 2008 to boost consumer spending in order to help jump start the economy.
This was accomplished through the issuance of asset backed securities.
The collateral for these securities was made up of auto loans, student loans, credit card loans, equipment loans, insurance premium finance loans, loans guaranteed by the Small Business Administration, residential mortgage servicing advances, or commercial mortgage loans.
Backing for these loans came from funds provided by the Federal Reserve Bank of New York.
Under the TALF, the Federal Reserve announced that it would lend up to $1 trillion USD in loans on a non-recourse basis to holders of qualified AAA-rated ABS backed by new and legacy securities.
The Federal Reserve extended loans in an amount equal to the market value of the ABS, less a haircut, and these loans were secured at all times by the ABS itself.
In simple words, you could buy certain securities with up to 95% of borrowed money from the Federal Reserve at very low interest rates, and in the case that the security did not perform as expected or did not generate enough cash to pay interest on the loan granted by the Federal Reserve, you owe nothing to them.
The TALF opportunity was unique since financial & regulatory constraints in the US initially did not allow investment banks (such as Goldman Sachs and others) to actively participate in this opportunity.
The opportunity was equal to none for high net worth individuals and a few institutional investors not bound by government constraints, however, very few moved fast enough to seize the opportunity, since these types of instruments were not well understood by the bulk of the financial community.
TALF enabled above-average returns to those with the know-how and the capital to invest in certain types of securities, in an artificially created low- risk environment.
To understand TALF, we need to understand first what an asset backed security is and position ourselves in the period 2007-2009, when macro events prior to the financial crisis led to the crisis itself (At that time, I was actively trying to raise capital to take advantage of said opportunity.
)What is an Asset Backed Securities (ABS)?An ABS is a widely accepted form of structured debt that is repaid solely by cash flows from a pool of isolated assets.
While the securitization market has traditionally embraced consumer and commercial assets such as auto loans, credit card receivables, mortgages and equipment leases, the market has also accommodated esoteric risks such as project finance assets, natural catastrophes, man-made catastrophes (terrorist attacks), life settlements, and royalty and intellectual property assets.
In a basic securitization, assets are sold by an entity (“Originator”) to a bankruptcy-remote special-purpose vehicle (“SPV”), usually set up in an offshore jurisdiction.
The SPV then issues bonds into the capital markets, whose payment of principal and interest are tied to the performance of the assets.
The proceeds received through the bond issuance are passed to the Originator via the SPV (See figure 1 for an example of a hypothetical USD 1bn securitization).
Since assets are sold by the originator to the SPV, the ABS does not represent a liability on the originator’s balance sheet, and the transaction is recorded as a “true sale”.
Figure 1: A template for a $1bn UDS securitizationOnce the bonds have been retired, the remaining assets revert to the originator.
What’s the difference between a Corporate Bond and an Asset Backed Security?A corporate bond is a bond issue by a corporation.
Corporations borrow money in the bond market to expand operations or fund new business ventures, and they can be broadly classified in two categories: investment grade and speculative-grade (also known as high yield or “junk”) bonds.
Speculative-grade bonds are issued by companies perceived to have lower credit quality and higher default risk than more highly rated, investment grade companies.
The quality of the bond is rated by rating agencies, in a scale that generally ranges from “AAA” (high quality), to “D” (default).
The credit rating for a corporate bond is a function of the features of the company issuing the bonds, which is translated into the probability of the corporate to pay its obligations.
In a securitization, structured debt is repaid solely by cash flows from a pool of isolated assets.
The rating of the debt, and hence the interest rate associated with the debt, is based on the quality of the assets and the features designed by the structure.
An asset backed security (ABS) is a security whose income payments and hence value are derived from or backed by a specified pool of underlying assets.
The pool of assets is typically a group of small and illiquid assets which are unable to be sold individually.
Pooling the assets into financial instruments allows them to be sold to general investors, a process called securitization, and allows the risk of investing in the underlying assets to be diversified because each security will represent a fraction of the total value of the diverse pool of underlying assets.
The pools of underlying assets can include common payments from credit cards, auto loans, and mortgage loans, to esoteric cash flows from aircraft leases, royalty payments and movie revenues.
Amortizing ABS are generally sold and traded according to their average life rather than their stated maturity dates, as with corporate and government bonds.
Average life is the average length of time that each principal dollar in a pool is expected to be outstanding.
The credit rating for an ABS is a function of the features of the pool backing the bonds, which it is translated into the probability of the financial structure paying its obligations.
There are several decades of statistics reflecting the probabilities of a bond being upgraded to a higher rating during its life, as well as the probability of being downgraded.
This statistic is known as a credit migration probability.
However, not all types of securities have the same pattern of credit migration.
Figure 2: Observed 1 year credit migration for Asset “X”For corporate and ABS bonds, ratings can be downgraded if the credit quality of the issuer or pool deteriorates or upgraded if fundamentals improve.
An investment grade bond (AAA, AA, A, BBB), that migrates to a non investment grade (anything below BBB) is known as a “fallen angel”.
Figure 3: In September 2015, the US Environmental Protection Agency issued a notice of violation that Volkswagen (VW) had installed diesel engine software to circumvent the emission standards of the Clean Air Act.
ABS showed a high degree of resiliency in this specific case of risk for the VW.
Although there is certain correlation between the risk/return of a AAA rated corporate bond vs a AAA rated ABS with assets originated by the same corporate, this relationship can be disrupted at times, with the AAA corporate becoming more risky than the AAA ABS and vice versa, as the example in Figure 3 shows.
Back in the mid 2000s, front office quants on the sell side like me did not have to develop custom models to price a pool of assets for a client.
Nevertheless, we needed to comply with the models and guidelines set by “the smartest guys in the room”, the PhD quants developing the software and analytics that everybody was blindly using to price these instruments, and which were basically propagating hidden risks.
One of the smartest guys was David X.
Li, a Chinese quant “Rock Star”.
All of a sudden, he was involved in high finance, practically straight out of academia, and with very little domain expertise.
David held a Master’s in Economics, a Master’s in Actuarial Science, a Master’s in Business Administration (MBA), a Ph.
D.
in Statistics, aside from also being a programmer.
David X.
Li, PhD.
Had David been involved in the practical side of finance as opposed to the research side, he would have seen that his paper in the wrong hands had serious implications.
His model represented an over simplification of the world, and did not handle tail risk at all.
JPMorgan Chase, which practically had an obsession with hiring top quants from top universities, hired David Li to work in its risk analytics division.
While there, he published a paper in The Journal of Fixed Income titled “On Default Correlation: A Copula Function Approach.
” Using some relatively simple math and I would dare to say, naïve, Li came up with simplified way to model default correlation without even looking at historical default data.
His model and accompanying analytics became widely used due to its simplicity in its implementation.
Years after the crisis, David’s PhD mentor said: “It was a very simple mathematical answer almost anyone could use, and when you’ve got a hammer, everything suddenly looks like a nail”.
Wall Street jumped on it.
This equation killed many 401KsNevertheless, a few front office quants like me were not happy with this “canned” model that did not deal with tail risk, and we developed our own, compiling statistics that dated back to 1926 for corporate bonds, and ABS dating back to 1985, which at least in my case, fed custom made Markov Chain Monte Carlo simulations and provided answers beyond the standard models.
These custom models that took into account the real correlations among the assets in a given pool, as well as the behavior of each loan in a pool, painted a different picture about the risk of the bonds.
It turns out that certain AAA ABS were a lot riskier compared to what Li’s model was implying.
The few quants who understood this and were able to explain the opportunity (shorting the securities) to decision makers at hedge funds and actually got them to bet against some of these securities made tons of money.
(As most people know, there are a few Hollywood movies about this).
Others like me tried to raise capital, but were not successful since typically quants do not have trading track records and there was not an easy way to short the securities with the potential for the highest returns.
The collapse of Lehman Brothers and the wave of downgrades and defaults of billions after billions of securities rendered Li’s model useless.
Defaults that the model had not predicted piled up and wiped out trillions of dollars in investment.
As Li himself said of his own model years later: “The most dangerous part is when people believe everything coming out of it.
”Btw, to me, Google’s TensorFlow is starting to look like a hammer, at the same time that JPMorgan and others seem to be on hiring sprees again, this time hiring PhDs straight out of academia and with functional roles as Data Scientists.
If history is any indication, I believe it is prudent to remember Li’s story as a cautionary tale.
In order to keep this article a manageable size, and before the coding session, I will continue it in Part II, where I will cover:How large was the securitization market leading to the crisis?Who where the leaders in issuance?What was the impact of ABS on the US and global economies?How did the Federal Reserve jump start the economy with the TALF program?Positioning ourselves back in 2009, what were the expected returns of investing in ABS securities after the wave of defaults?Needless to say, I would love to hear from you with whatever Qs you might have up to this point.
.